Krishnakumar Gurumurthy argues that AI's frontier capabilities are outpacing business models built on human labor, creating a socio-economic gap disguised as a technical one. He proposes discretizing business processes into indivisible functional units, mapping them to appropriate AI techniques (neural operators, transformers, RL, etc.), and building a new enterprise stack with guardrails and trust architecture—drawing an extended analogy to how electricity safety standards (insulation, circuit breakers, licensed electricians) evolved to argue for similar AI containment, validation, and certification regimes.
Shared by the speakers for the symposium. Email addresses have been removed.
Transcript
YouTube’s automatic captions, unedited, so names and technical terms are often misheard. Each timestamp opens the video at that moment.
0:00 proliferation of machine intelligence converting business into hyper learning machines. So, by the way, my name is Krishna Kumar. And before we delve into the topic, I would like to thank the forum for the invite and also it’s a pleasure to and honor to present here in this protocol symposium. And a little bit about me, I have close to 15 or close to two decades of experience across multiple domains through e-commerce, health care, telecom, and automotive. And I’m basically an engineering in background and specialized in communication technology and my master’s is in applied neuroscience. So, basically, as a part of bachelor’s I was trained on layer one to layer seven of TCP OSI model. And as part of master’s it’s from layer eight to layer 10, the human part. So, now, let’s delve into the topic. Uh Yes, with these AI labs and these AI companies leading frontier models, they have launched these intelligent models or AI systems which has numerous capabilities and uh uh all potentials, that’s extremely good, extremely useful, but the business models are rooted on human expertise, labor, and skill. So, obviously, there is a gap. And this gap, on the surface level, it looks like a technical gap or a tooling deficit. So, we try to use those AI systems as a charge I mean, charge GPT, Claude, and all those things and then we build some kind of a bot.
1:30 And yes, we try to redesign our organizations and enterprise systems around it. Yes, but it’s not just purely a technical gap, it’s also a socio-economic problem in a technical clothing, basically. So, the entire presentation is going to deal with this, not completely on the socio-economic part, but see what we can control, how we can navigate and uh yeah, take it forward. So, basically, before we even get into the problem, we need to uh understand that this functional space of the business domain needs to be first discretized, basically split into an indivisible units. And then you need to select the technique, the AI technique based on the dimensionality of the domain. And then the process needs to be redesigned, which involves little bit of an organizational redesign that in turn leads to an socio-economic settlement. So, that’s the proposition that I start with uh basically the gap. Now, the thesis argument is basically I would like to split into four parts. First is to discretize the functional space, the business domain.
2:39 Uh number two, it’s a regime change. Basically, as I mentioned in my initial uh slide, it’s a system of records as in the traditional software that we do. It’s basically we maintain the records, basically, to a system of inference. All these agents. So, that’s uh paradigm change or regime change, that’s what I call about call. And then there is a lot of labor reskilling and labor displacement around this are the type of labor it will going to change. What is scarce is going to be become cheap. I mean, it’s going to be too much and uh there’s going to be a new set of winners and losers, basically, in terms of labor market. And then the architecture pathway that’s going to drive this transition and also the necess- and also the final target architecture, how it might look like just that. And discretization, let’s start with. So, basically, as I mentioned earlier, any domain, a business domain or a functional space, can be decomposed into discrete label functions. Uh it’s an indivisible units. And then those domains, those indivisible units, will act as a transformation part.
3:41 Basically, you have some kind of a goods and services or capital and that gets transformed into something and you get an output. So, the transformation efficiency is called as profitability, of course, which is relative to the market in which where this domain is operating. So, but still if you look at it, the functional skeleton of this transformation structure is um invariant across different domains. Or you could trade off with some little bit of reconfigurability. And this entire discretization process of a functional domain can be viewed as in a geometrical structures. Basically, all your input can be called as an entry points and your transformation will compress a lot of intervention points. Uh Emerging characters will Emerging uh phenomenon will evolve there. And then the extraction points, which is your output, where you measure your values. And these uh Here I would like to quote Donella Meadows, uh what do you call the leverage of on leverage points? They say 14 stacks of uh leverage points. They say the lowest stack of leverage point talks about parameter change to control the system and the highest level of leverage point is about paradigm shift.
4:53 So, if we organize these interaction I mean intervention points or interaction points in a the hierarchy of abstraction principle, yes, could do a good job there in the transformation part. And this transformation should provide adaptiveness with stability. Basically, enough plasticity to reconfigure under different environment or different task and without losing the structural integrity. In case if there is any reconfiguration, it shouldn’t dissolve. So, that is about the geometrical part of this. And once this is done, then that then we need to apply intelligence to exactly where the decision is made. So, it becomes either an autonomous or a semi-autonomous or human oversight or human in the loop. So, these are the five parts in this entire domain, which is basically first is about effect transformation where we change uh a good from one form to another or one function to another a service in one function to another and then the process orchestration we move around the the the the domain in terms of workflow process and dependencies and basically gaps between what is intended and what is actual so basically so process orchestration and then finally the value capture.
6:11 So your goods and services moves in one direction and the economic value moves in the opposite direction to the one the the guy who’s providing it. So basically we need to capture the value that should align with the operational reality and the captured value should be a we should be able to verify it retroactively and that’s what you call it as assurance and auditability. So the operational reality doesn’t stay true during that time rather in hindsight it should be the integrity has to be maintained and sustained for decades. So that’s what assurance and auditability comes and the life cycle traceability for any goods and services and capital all the way the life cycle from the beginning from the concept stage to a reality is delivered or becomes extinct needs to be completely documented and it should be traceable so that in case of any failure this entire process we should be able to trace it back to the origin of the failure to understand and do a postmortem and then make sure that it doesn’t happen again. And any decision that we talk about autonomous decisions or agentic decisions that have more or less three states. The first one is it decides it needs to have an observable set of signals on which it needs to operate. And then it needs to make a decision the choice on what what direction make a recommendation or an action whatever it is. And then after the choice is made that needs to be recorded and then needs to be evaluated against the evidence that gives a quality of the decisions.
7:36 So usually the third part is the challenging part. How do we implement the third part defines whether the system is going to succeed or not because the failure here is not obvious.
7:47 The failure is undetected, invisible. So, it will cause systemic drift when unchecked for it could be for years or years and finally it will lead to a system collapse or irreversible failure, I would say so. And I talked about the two regimes that I talked about. 1.0 mean I would like to borrow this term 1.0 from Andrej Karpathy, the famous guy I I guy. So, next generation of softwares. I don’t want to do 2.0, 3.0. I don’t know what what version it’s going to evolve. So, at a high level we have compared against these kind of dimensions like knowledge local semantics, retrieval, computation, competence, boundary, failure mode, assurance basis, locus of inference. So, at a high level your 1.0 is explicit memory whereas the next generation is an implicit memory represented using ontologies and um representational memory. We’ll talk about ontology and representational memory in detail in the subsequent slides. In the first one it is more of the inference is done by the operator. Here operator is your system user or a business user.
8:54 Whereas in next generation, the inference is internalized. So, that’s where the challenges comes from. It gives a tremendous capability, it reduces the cost down, but it brings other set of challenges. So, go to the next one or you want to take a look. Take a look. So, the first regime is more of an explicit memory and set of rules that is authored by operators. And it’s deterministic, the provenance is transparent, but the only problem is the combinatorial brittleness which is basically when you have humongous amount of rules, uh interlocking rules, you could end up the systems could degrade and continuously under lot of boundary conditions could fail or you need to throw in huge amount of human labor which means it will increase the cost of maintenance. I mean it’s basically it’s going to be exponentially high basically. And that’s the regime 1.0 which we have been practicing for decades.
10:02 And regime 2.0 or the next regime which form of the memory is kind of internalized. So ontology is one form of memory where we are this more or less like your vector embeddings that we talk about today all the vector database where we it’s all the structural correctness and constraints are enforced in case of ontologies. When there is change in ontology which is basically in organizational reality or in the market reality if something changes you can go and edit it and keep it track. So you have this option. Whereas the representational memories are nothing but your weights and biases of the model. So that gets trained with the data. And of course there is a lot of security aspects of it how to maintain the integrity that we can talk about a little bit broadly these are two types of internal memory in regime 2.0 but if you drill down there is more types of memories we can bring in but that itself is a separate topic the memory itself because there are there will be a lot of different kinds of memory that we might need when we evolve into the system next generation. And operating conditions so basically the manifestation of these AI systems could be any of these six models six types assessment where it gives a description for anything a shop or a concept or anything and then it gives an evaluation you you submit a slide or you submit a a report gives evaluates it against a target domain under it’s a recommendation travel recommendation generation prediction and action. But, all the six ways of manifesting a system is should be carefully coupled with something called feedback, which is a learning signal.
11:41 The cycle closes in a feedback. So, it’s basically the feedback is a critical part here because if lack of clear feedback signal or a poison feedback signal or if it is not rooted in say hardware root of trust or something quite concrete, it could cause system to fail invisibly. And it gives and uh the what do you call the technical terminology, they call it epistemic drift, which will go undetected for ages. And when it gets detected, the blast radius is so high and the impact is irreversible. And that is where now coming to the architecture part, uh the existing one is we call it enterprise architecture. For all the organization, you’ll have an enterprise architecture. So, all these enterprise architectures are based on frameworks. There’s plenty of frameworks. TOGAF, ArchiMate, PRINCE2, a lot of enterprise architecture framework where it gives an authoritative answer for anything and everything how you want to do in an organization. Yes. But, when we move towards these new kind of AI systems, what we need is a little bit of uh more high discipline, uh which is rooted in uh some kind of uh disciplines I mean complex adaptive systems. So, I would like to touch upon either mechanical engineering, aerospace, or electrical engineering which follow system engineering principles. So, if we could take some lessons out of it. So, that’s where we call quantified constant. So, basically traceable requirements, calculated error margins, and failure modes are enumerated, and verification and validations are taken care.
13:20 So, but this borrowing of these system engineering from these other disciplines into this enterprise system engineering doesn’t come with some it comes with some caveats. What is the caveat? Basically, the system engineering principles are rooted on physical laws. Be it aerospace engineering, mechanical engineering, or thermodynamics laws or Bernoulli’s laws all those things. Whereas in case of enterprise software engineering, there is nothing of a physical laws. There’s nothing that’s going to keep them rooted or grounded. So, we need something so concrete that or else it’s just borrowing the terminologies will not guarantee the substance that is required to dive the system duty. So, that’s where the protocol infrastructure kind of thing can help. Um need a dumb set of standards, not just in terms of verbal standards, verbal form, but it needs to be in reality enforceable at a ground level.
14:12 And uh my view or my take on how this enterprise system the new enterprise system stack could look like. It’s basically intent driven interface, which is your chat interfaces, and which is backed by your autonomous multi-agent core execution platform, which in turn is underneath you have your self-synthesizing logics and continuous semantic injection, which is injection, which is basically uh your ontological ontological models are renewed on a day on a daily or on a near real time or on a real time basis. So, which aligns with the organizational reality and the market reality. And all these four layers produces enormous amount of features. It cuts down your maintenance cost. And it shifts your human labor. But all these layers four layers produces something, whereas the last layer, which is your guardrails and trust architecture, doesn’t produce, rather it acts as a constraint. Which is uh basically it’s a negative producer, I would say so. But that is essential. These are features that is the basic first principle on which you need to start the rest of the layers. So, that’s uh the gravity, I would say so.
15:16 And the principle the guiding principles for all the four layers reliability maintainability, which even today we do it, which includes security, all those things. But on top of it, even today we take care of safety and human factors, software system, but not to a greater level. So, more emphasis required, especially on the dimensions of AI safety, when you talk about there is different dimensions like psychological AI safety and many human dimensions comes into picture. That’s what I talk about beyond layer seven comes into this picture, not just TCP/IP protocol or OSI layers that you talk about. You need to secure the application layer, presentation layer or transport layer, but we need to secure layers beyond seven. So, that’s where this AI, I mean, the safety and the human factors comes into picture. And finally, the guardrails in the trust architecture. So, this requires a separate slide because this guardrails and trust architectures roots is rooted on four pillars. This is your aligned containment, validation, and runtime assurance. So, systems optimization should be aligned with the organizational actual objectives and the reality. So, when there is a misalignment and that needs to be detected and need to feed back into system to the corrective action.
16:29 And runtime assurance, and traditionally we have this QA, I mean, quality assurance. So, you have different levels of testing before the software is shipped to production. Now, that doesn’t guarantee you a good successful run. Rather, when the software or a model or something is deployed in production, you need to ensure that the model behaves as it should. So, we need different set of monitoring the runtime behavior of the systems. And with this, I mean, that’s what we talk about different guardrails. So, and containment, basically in case if any issue happen, you need to restrict the blast radius. Let’s say the the hugging face incident, that’s a very good example of violating the containment. And validation, which is a demonstrable, I mean, it’s basically validation and verification. So, your grounding technologies, basically when you have a behavior of a model that’s being deployed and that needs to align with the ground truth on a regular basis. So, you need to have validation and verification mechanism to ensure it. And now, I mean, of course, this is my interesting perspective.
17:34 Some 100 years ago, we invented electricity. So, when electricity was invented, we had we talked about human safety. Maybe a couple of dogs and humans might have died due to electrocution. So, we kind of came up with a lot of things over a period of time called insulation, circuit breakers, grounding techniques, some building codes on how the electrical wiring needs to be done and some certification labs on certain electrical equipments and licensed electrician. Not everyone can go and do the electrical wiring in the house. So, similar to that, in AI also, containment is I like to draw this parallels here. So, the containment is same as your insulation and circuit breakers is your monitoring rate limits. Even today in software 1.0, we do have something equivalent of circuit breakers in today’s domain, but we need more advanced version of it. And grounding technique, which is basically your human in the loop, RL, RHLF. And RLHF and human oversight and appeal for high-stake decisions. You cannot make it autonomous, needs to be semi-autonomous or human in the loop. And build all your AI standards that comes as a building codes here. Can compare it to building codes.
18:45 And certification labs. So, here this is where we need some kind of a red teaming or maybe third-party auditing companies that regularly audit the model before that’s being deployed for the production or the public use. I think there is a lot of talk noise going noise or scream going on in a lot of international safety association forum about forming a panel to ensure the models, especially the frontier model that’s deployed to the public, needs to go through certain basic level of sanity and the certification before opening it up for the world. And then, licensed electrician is equivalent to a builder. So, the builders who are responsible for deploying these AI models needs to clearly understand the liability. I mean, liability and the responsibility that rests on them. Cuz earlier days, a developer makes a small defect that could uh cause maybe a few failures, you could roll it back, and you could trace back, and you do a postmortem. And now, it’s yeah. I As I told you, the failure in most of the cases goes undetected.
19:48 And public literacy is also quite important. And but of course, this analogy has certain drawbacks. I mean, there are breakpoints like the electricity obeys physics. Whereas AI systems, I don’t know there is any law. So, basically, your alignment and interpretability interpretability research is more of like a Ohm’s law, but that’s not yet fully developed. And that’s passive, and this one is an active one. And deliberate misuse. You have AI cyber attacks, whereas electricity insulators could protect you. But Yeah, and if someone wants to electrocute other person, then there is no blocker There is no prevention, right? So, whereas in AI, this deliberate misuse is possible. And the pace is it’s ways apart in case of electricity and AI. [clears throat] And global reach as well. So, this is another interesting paper that we’d like to cite. It’s a little away from the regular discussion where this paper discuss about a knowledge collapse. Uh this is from AI human cognition and knowledge collapse. It’s an MIT paper published in Feb 2026, where this talks about very important thing. You have a general knowledge base, and you have a specialized knowledge base. So, if a tax consultant who has three decades of experience, and he has with all his general knowledge base, he can suggest a company on her tax evasion and tax recommendations, and it works. And finally, an agent AI that trained on the general knowledge base and if a new pass out a newbie who joins the workforce and he just uses agent AI to give a similar or more or less same as the expert has given. So what happens is in a short term there is a huge development in terms of capability. The companies could make a lot of profit but in a long run the giving back to the society which means giving back to the general knowledge pool stops. So even today the stack overflow traffic has dropped. Wikipedia traffic has dropped. So what it happens is in long run the knowledge collapse is imminent if unless otherwise we incentivize the learning and the research and curiosity and contribution back to the society. The knowledge base is important. So we need to design as policies and systems to keep the human effort flows in optimal.
21:56 And of course all these things require cross domain borrowing but software engineering in its native form doesn’t have these complex adaptive behaviors. So we need to borrow some terminologies from other disciplines. So there’s one example I would say entropy. Entropy in information theory is a vocabulary borrowed from thermodynamics. It’s a very good example. It was borrowed not just as a word but as a substance. So there are some principles that guiding principles that can allow this borrowing to be very effective because the structural condition has to be comparable and the prediction should be working under the prediction should be falsifiable. So these are some of the domains disciplines that I have enlisted and some of the terminologies that I have enlisted that would could help but it could be more of this or less of this. And finally I would like to conclusion. Sorry. So as a part of conclusion I would like to emphasize more on the safety. So we have enough amount of capabilities gained as a part of this new AI systems AI labs and frontier models but the visibility of failure is lost.
23:02 Uh that’s a conclusion note that I would like to say. But when you say the visibility of failure is lost, so what’s the consequence? Which is basically your capabilities have been concentrated into fewer practitioners. Eventually, when we adopt these things. So when when fewer practitioners has huge amount of capabilities in their hand, what happens is there are two options. One is the human capacity increases, expands, and uh our part of the human labor can be displaced into something else. But if we design the systems transitioning with the responsible architecture, we could avoid failures, silent failures, which are dangerous one. Then doesn’t matter whether the human labor is displaced or expands.
23:52 Uh it doesn’t matter. Whereas if we don’t design the systems in a responsible way, it could encounter these kind of failures at a later stage or catastrophic failures or it could lead to a polycrisis or something like that. Then the difference between human capacity expansion and the displacement matters, and the displacement is a serious issue. It’s a serious social issue. That in itself is a separate topic to be discussed, to be researched as well. And all these systems that I talked about, software systems, they are not just merely software. They are they carry the economic activity. Basically, it’s same as your railroads and roadways carries uh passenger and transportation load, right? So it’s an infrastructure carries a load. So when you are carrying a huge amount of economic activity, the basically the discipline of its design and the constraints before capabilities, failure modes, these needs to be carefully take carefully addressed the issues. Uh basically, safe by design or yeah, not safe at all.
24:54 So yeah, that’s my concluding note because uh this entire enterprise all these new things gives lot of excitement and adoption trying of new things but that has a lot of caveats and some of the technologies required to address these kind of caveats are not yet productionized I would say so. And also the the capabilities are also keep on increasing these days the pace in which they increase because the more and more the model accuracy increases it’s going to create new set of winners and losers. So in the paper that I suggested one of the authors suggestion the paper that I’m talking about this one the authors suggestion is two things to temporarily ban the this cloud or open AI and to reduce the accuracy of the models. Well, that’s a controversial claim I would say so but there are other approaches also but yeah, it’s subject to the people who read the paper.
25:57 Yeah, with that I would like to conclude the talk and if there are any questions I’ll be happy to take it now.
26:08 I have a quick question Krishna do you go by Krishna or Krishna Kumar?
26:15 All right. Yeah. We talked earlier and I’ve been thinking about this since then I haven’t been in the enterprise environment for almost 15 years I’ve almost lost touch but one of the things I get whenever I see people on LinkedIn posting about AI or even talks like this immediate question that comes to mind is is anybody having fun in enterprise IT with AI or is it just one big headache because when I see like normal people playing with AI tools it seems like they’re having fun they’re making images writing like coding stuff but when I talk or look at enterprise IT people, it doesn’t feel like anybody’s having fun.
26:55 No, I don’t think so because it’s at a very rudimentary level. A lot of people are using it for white coding, code reviews, and things like that. When you say people are using AI even in my company, what they do is they are using AI for code reviews or maybe some of the codes are white coding and things like that, but the real damage, I’d say, when it gets into the real production situations, safety critical systems, the I don’t think people understand the consequence of the failures there. Basically, when you productionize these systems into when it impacts real life, let’s say autonomous systems, say self-driving cars. So, in all those areas, I don’t think the consequence is not really They talk They mostly they talk about jobs.
27:39 What happens if the job goes away? AI takes away a job. That’s the major discussion that dominates in the topic of AI in enterprise, but it’s very sad, I would say. So, So, Any more questions?
28:00 Yeah, I think that was it. Thank you so much for the talk. And then we’ve got one more before we go back to two tracks. So, thank you again, Krishnakumar, for the talk and looking forward to seeing you around on the Discord.
28:15 Yep. Thank you. Thanks for the invite and opportunity.
Amber Case presents the Calm Technology Institute's 81-point certification framework, arguing that human attention operates like a law of nature—distinguishing prefrontal-cortex-demanding 'center' attention from pass-through 'peripheral' attention—and that products (from car buttons to AI interfaces) should be designed and governed accordingly. She traces the framework's lineage to Mark Weiser's 1995 calm technology principles and discusses the legitimacy problem of building certification authority without statutory power, comparing CTI to Underwriter Laboratories and the Good Housekeeping Seal. The session closes with audience discussion, including Venkatesh Rao connecting the talk to Kairos/Chronos time and AI's 'staircase' temporality.
Daniel Kronovet presents 'Chores as Complex Coordination,' introducing 'regenerative labor' as a concept distinct from maintenance or care work, and detailing Chore Wheel, a Slack-based protocol suite (chores, hearts, things) built on Elinor Ostrom's commons-governance principles plus four digital-era additions: no privileged roles, simple inputs, humans-for-sensing/machines-for-bookkeeping, and continuous asynchronous participation. Drawing on his Los Angeles co-living house Sage House as a real-world case study, he argues that separating task definition from task performance (akin to double-entry bookkeeping) allows a fixed, continuously-replenished point economy to turn dish-washing and trash duty into a system of social computation that generates social capital alongside physical upkeep.
Nicholas Fett traces the history of market neutrality from 1889 Oklahoma land runs and Adam Smith's moral sentiments through Keynes, the Mont Pelerin Society, and into crypto, arguing that protocols designed to be neutral (Bitcoin's decentralization, Ethereum's 'crops' ethos) inevitably get captured by economic incentives like mining and block-builder centralization. He uses the 'tunnel' and 'one more lane' analogies to argue that building for neutral principles without a defined societal goal leads to outcomes divorced from original intentions, and that protocol designers must instead encode explicit purpose rather than chase an unattainable neutral equilibrium.